Recently, Professor Shen Dinggang, MICCAI Fellow, Founding Dean of the School of Biomedical Engineering, ShanghaiTech University, and Co-CEO of United Imaging Intelligence, made a presentation entitled “Application and Innovation Research of Artificial Intelligence in Full-Stack and Full-spectrum Medical Imaging” at the 25th MICCAI Conference. “(Full-stack Full-spectrum AI in Medical Imaging), based on the practical experience of medical AI in China, put forward a new idea of full-stack full-spectrum medical imaging innovation that connects industry and academia.
The MICCAI Society is an internationally recognized top comprehensive academic organization in the fields of Medical Imaging Computing (MIC) and Computer Assisted Intervention (CAI).
The content of the conference reflects the latest achievements in the field of medical imaging, represents a new direction for the development of the discipline, and has an internationally recognized academic status and influence.
This year’s MICCAI conference was held in Singapore from September 18 to 22. Professor Shen Dinggang was one of the three invited keynote speakers at the conference.
After the meeting, Leifeng.com had a dialogue with Professor Shen Dinggang, provided many valuable opinions on changing research thinking and breaking the interests of circles, and shared the latest construction results of the company and the college.
He said that the three keynote speeches of MICCAI will include two directions of MIC and CAI, and the other will discuss more macro topics, such as AI ethics and fairness this year.
The three keynote speeches will be recommended by scholars in different directions, then a meeting will be held to determine, and finally a special chairman (different from the chairman of the conference procedure) will be invited.
The scholar who invited Professor Shen Dinggang was Professor Pierre Jannin of the University of Rennes, France.
He has approximately 20 years of experience in designing and developing image-guided surgical systems for neurosurgery, is Secretary General of the International Society for Computer-Assisted Surgery (ISCAS), and serves as an associate editor and reviewer for several journals.
In 2018, Professor Pierre Jannin was elected as a MICCAI Fellow.
According to Leifeng.com, the speech was prepared for nearly 3 months, and a team of more than ten people including Shanghai University of Science and Technology and United Imaging Intelligence participated in the speech. The key word of the speech topic is “full stack and full spectrum”, which contains Professor Shen Dinggang’s far-reaching thinking on the development of the discipline.
He said that many academic studies are only used in a certain part of medical imaging sporadically, and will not bring significant improvement to the clinical workflow as a whole. Only by combining AI methods from a full-stack and full-spectrum perspective, and researching and developing medical imaging AI from the whole process of pre-scanning, scanning, and post-scanning, can it truly bring about innovation in clinical workflow.
“For example, the diagnosis and treatment process of a disease is a rope. You have done a good job in image segmentation and classification at the front end, adding 5% upward force to the rope, but the force transmitted to the end of the rope has been lost. So, if you only focus on one of the points, the probability of moving the whole rope in the end is very low.”
Therefore, Professor Shen Dinggang hopes that scholars will use the thinking of “points into lines” to consider problems, stack up many full-stack and full-spectrum “lines” to form “surfaces”, and solve real clinical problems, that is, “point-line- face” idea.
“In the era of AI, we should consider the overall solution of a problem, rather than just one point. Some people have drilled deeply and drilled for a lifetime, but looking back, what is the benefit of this in solving clinical problems. It can be said Most of them are of little benefit and use. Because what is needed clinically is a large number of AI tools.”
How to do this, the key is to standardize and modularize all AI tools, combine them, and quickly develop new products. This requires scholars and industry people to learn to share and start from the idea of solving clinical problems.
For young scholars, Professor Shen Dinggang suggested that they go to large comprehensive laboratories and learn to look at problems from a global perspective.
“Students who do image segmentation should think about how their work can help and support each other in their work, so that they can get comprehensive training.”
Professor Shen Dinggang was the chairman of MICCAI 2019. Under his impetus, the influence of the Chinese on the world academic stage has grown.
In terms of papers, in 2019, the proportion of accepted articles in Asia has reached 37% (36% in the Americas and 26% in Europe), surpassing that in the Americas, and the vast majority (about 150) of these accepted articles came from China.
He did a few things to expand the number of reviewers to around 1300, and implemented a very good strategy:
First, let each of the more than 60 field chairmen of MICCAI 2019 recommend at least a certain number of experts;
Secondly, in MICCAI, the number of Chinese field chairmen is usually very small. One change made by Professor Shen Dinggang is to increase the proportion of Chinese holding field chairpersons, from the original 13.8% (MICCAI 2018) to 43.5% this year. Thirty field chairmen are Chinese.
Therefore, a great contribution of MICCAI to the domestic medical AI field is that it not only promotes young scholars as field chairmen to the world, but also promotes China’s work in medical imaging AI to the world.
However, with the increasing number of the best papers of the conference and the presence of Chinese young scholars (for the list of reviewers of this conference, please refer to the original text), it also aroused the concern of Professor Shen Dinggang.
He thinks this is not a good phenomenon.
Because scholars from a certain country or region occupy an important position, they will form small cultural circles and interest circles, and the thesis will appear opportunistic or “personal points”.
In some extreme cases, the double-blind mechanism of MICCAI review will also fail, which is something he does not want to see.
Professor Shen Dinggang said, “I hope that people in China can do better and better in the field of MICCAI, and I hope that everyone can support each other and give more opportunities to young people, but this should not be a springboard for academic profit, but to maintain a pure The original intention of academics is to do practical things for medicine and patients.”
Currently, Professor Shen Dinggang is working at ShanghaiTech University and United Imaging Intelligence, and he shared his progress.
In August, the first batch of major projects of the National Science and Technology Innovation 2030-“Brain Science and Brain-like Research” were successively opened.
Among them, the “Infant Brain Development Cohort” major project and the “ShanghaiTech University-Child and Adolescent Development Cohort Construction Sub-project” under the guidance of Professor Shen Dinggang, the founding dean of the School of Biomedical Engineering, ShanghaiTech University, were officially launched recently.
The “Infant Brain Development Cohort” project is led by Zhang Han, a researcher, associate professor, doctoral supervisor, and director of the Laboratory of Infant Brain Development Imaging, School of Biomedical Engineering, ShanghaiTech University;
“ShanghaiTech University-Child and Adolescent Development Cohort Construction Sub-project” is led by Wang Gan, a researcher, associate professor, doctoral supervisor and assistant dean of the School of Biomedical Engineering, ShanghaiTech University.
The project is benchmarked against the highest international level, and will be based on United Imaging’s latest scientific research 3.0T magnetic resonance system to conduct research on brain development covering two important age groups, infants and children and adolescents.
Professor Shen Dinggang called it a “historic” phenomenon.
“The magnetic resonance machine that is widely used in brain science research is driven by the brain project research in the United States. Therefore, in major national frontier scientific research projects, it is necessary to study and solve many scientific problems in the field of brain science, and it should also drive domestic The development of related technologies and industries has given high-end imaging equipment companies more opportunities to polish.”
In the two projects/projects, ShanghaiTech University will cooperate closely with the Pediatric Hospital of Fudan University, Shanghai Normal University, Xi’an Jiaotong University, Shanghai University and other institutions.
In addition to being the founding dean of the School of Biomedical Engineering, ShanghaiTech University, Professor Shen Dinggang is also the co-CEO of United Imaging Intelligence.
He said that at present, United Imaging has more than 40 products, and the products have been deployed in thousands of hospitals, with sales of over 100 million yuan.
The following is an excerpt of Professor Shen Dinggang’s speech at the conference.
Faced with complex and diverse clinical scenarios, a full-stack and full-spectrum examination is required
In recent years, medical AI has ushered in great development on a global scale. More and more top scholars in the industry have devoted themselves to the field of medical imaging AI. Since 2018, the number of medical AI-related articles received and published by MICCAI has also increased significantly. This can be seen from the keywords that MICCAI is frequently researched.
Frequent words in MICCAI articles
However, many academic studies are only sporadically used in a certain link of medical imaging, and will not bring significant improvement to the clinical workflow as a whole.
Only by combining AI methods from a full-stack and full-spectrum perspective, and researching and developing medical imaging AI from the whole process of pre-scanning, scanning, and post-scanning, can it truly bring about innovation in clinical workflow.
In addition, many algorithms can only identify a single lesion or disease from the image, but in the clinic, doctors need to examine all possible diseases at the same time and issue a report.
On the other hand, clinical diagnosis often uses not only one modality, but a combination of multiple modalities to complete the examination. Therefore, we need to study how to diagnose diseases from a full-spectrum (i.e., multimodal) perspective.
Aiming at complex and diverse clinical problems, in industrial practice, we conduct full-stack (disease diagnosis and treatment workflows that run through imaging, screening, follow-up, diagnosis, and treatment) and full-spectrum (covering X- ray, CT, MR, PET, PET-CT and other multi-modalities) medical imaging AI research, and developed a general full-stack full-spectrum technology module. Like Rubik’s Cube components, these general modules can quickly and efficiently form different products and solutions through different arrangements and combinations, bringing practical value to the clinic.
Full-stack AI, running through the whole process of diagnosis and treatment
Taking lung cancer as an example, there are 2.2 million new cases of lung cancer and 1.8 million deaths due to lung cancer every year. Lung cancer diagnosis and treatment also faces pain points such as small screening coverage, low efficiency, many lost patients during follow-up, and difficult diagnosis and treatment. In response to these pain points, we have built a full-process full-stack AI for intelligent health management of lung cancer, including functions such as health assessment, screening, diagnosis, treatment, prognosis, and follow-up. AI empowers the clinic to bring good news to lung cancer patients.
Lung cancer intelligent health management full-stack AI first focuses on low-dose CT reconstruction, reducing radiation dose by 30%-70%, improving imaging quality, and helping doctors to improve the detection rate of pulmonary nodules by 8%-15%. FDA certified.
Based on low-dose CT images, a multi-task and multi-scale deep neural network is constructed to realize the segmentation and labeling of trachea and lung parenchyma, and to locate, classify and segment nodules, and realize lung nodule screening and diagnosis. This product has been approved by China. The NMPA Class III certificate has been applied to more than 800 hospitals in China, and has achieved daily screening of more than 50,000 patients.
During the follow-up process, multiple images of follow-up patients at different times are matched to anatomical structures and lesions through the registration engine, and nodule changes are automatically quantified. The follow-up system has been used in more than 600 hospitals, with a monthly usage of more than 1.5 million times.
In radiotherapy, United Imaging’s intelligent segmentation engine can automatically delineate the target area in about 0.7 seconds, and the Dice coefficient can reach 97%. The integrated CT-linac radiotherapy equipment equipped with an automatic delineation engine shortens the traditional first radiotherapy process that takes about 20 days to complete to a one-stop radiotherapy of about 20 minutes, “precisely targeting” lung cancer lesions.
One-stop radiotherapy has been successfully applied to rectal cancer (23mins), nasopharyngeal cancer (29mins), breast cancer (23mins) and lung cancer online adaptation in Fudan University Affiliated Cancer Hospital, Sun Yat-sen University Cancer Center, Jinhua Central Hospital (planned adjustment 10 mins 45 sec) in the first radiotherapy of the patient.
Full-spectrum AI, enabling multi-modal devices
In the fast low-dose scanning of different modalities, the blurred, artifact or under-sampled images can be restored to high-definition images equivalent to full-sampling quality based on image multi-domain mapping technology. These techniques can be applied to reconstruction tasks of CT, MRI, PET and other images.
For example, the AI-assisted Compressed Sensing technology ACS (AI-assisted Compressed Sensing) applied to MRI, after training with millions of data, can complete multi-sequence imaging of various parts of the body within 100 seconds, not only maintaining high-definition image quality, but also greatly Shorten MRI scan time. This ACS technology is FDA-cleared and has been used on more than 250 MRI machines.
Image multi-domain mapping can also be applied to low-dose PET reconstruction. The HYPER-DLR (HYPER Deep Learning-based Reconstruction) technology developed by United Imaging can improve the scanning speed by nearly 10 times with the same image quality. The technology has enabled more than 250 PET machines.
ACS technology and HYPER-DLR technology
Rubik’s cube combination, infinite transformation
An effective AI modular engine can effectively enable full-stack workflow and full-spectrum device modalities. Taking the segmentation engine of United Imaging Intelligence as an example, it can achieve the goal of accurate segmentation in seconds on multi-modal images. By developing artificial intelligence technology modules such as detection, segmentation, registration, classification, and mapping, different clinical products can be developed in a short period of time through different arrangements and combinations.
For example, during the COVID-19 period, we quickly developed a new coronary pneumonia AI product that runs through the entire process of automatic scanning, diagnosis, quantification, and follow-up within two weeks, and implemented it in more than 320 hospitals to help doctors quickly scan and diagnose;
We developed brain metastases detection software from scratch in 1 month. After multi-center clinical verification, on the basis of no significant change in false positives, assisted doctors improved the detection sensitivity by 21%, and the average reading of each case was obtained. Time reduced by 40%;
In the auxiliary diagnosis of stroke, the development of four major AI systems in the one-stop stroke solution (hemorrhage, ischemia, CTP, CTA) has been completed.
Rapidly developed products based on common modules
Industry-University-Medical Collaboration to Promote the Development of Frontier Technology
The rapid transformation of AI applications from research and development to implementation is inseparable from the accumulation and combination of technical modules, as well as the close collaboration between industry, academia and medicine.
In the field of brain science, we are conducting research on brain health throughout the life cycle of 0-80 years old through industry-university-medical collaboration, including early brain development from 0-6 years old, brain aging research, and MR-based minute-level brain health screening.
In terms of early brain development research, ShanghaiTech University has taken the lead in undertaking the “Infant Brain Development Cohort” major project, which is an important part of the China Brain Project.
In this project, ShanghaiTech University cooperated with a number of universities, hospitals and enterprises to introduce United Imaging’s high-end imaging equipment and United Imaging’s intelligent AI technology to jointly conduct research on early brain development of 0-6 years old.
In terms of data collection, it breaks through the traditional calm scanning method. By letting infants and young children familiarize themselves with the venue and scan in sleep state, the impact of scanning can be minimized.
Considering that infants and young children are particularly sensitive to the noise and scanning time of MRI, ACS technology can achieve fast scanning, save 44% of time, and greatly improve the quality and success rate of data acquisition;
Through MTP technology, 16 contrasts and 9 quantitative maps can be generated in one acquisition, and rich measurement sequences can be obtained;
In order to avoid poor scan results caused by the baby’s body movement, the wearable head monitor can monitor the baby’s head movement during the scan and adjust the gradient prospectively to perform motion compensation to make the scan clearer;
We also custom-made baby coils for a comfortable scanning experience for babies.
In the entire study, the technology we used has been applied to United Imaging’s scientific research-grade high-end magnetic resonance uMR 890, which has also been installed in the School of Biomedical Engineering, ShanghaiTech University, and plans to scan 1,000 research subjects in the next five years.
With the growth and development of infants and young children, using different modalities of examination in different time dimensions and combining different technical analysis methods, a variety of research results can be obtained, forming a full-stack and full-spectrum brain development image across the developmental cycle. Research, and then use professional scales to assess the development of children, and issue a brain development health report, which can fully grasp the growth and development of the brain.
In terms of brain aging research, we can perform cortical segmentation, Alzheimer’s risk monitoring, follow-up, hippocampal atrophy assessment, cerebral small vessel disease assessment, etc. through AI, and provide detailed brain health assessment reports.
In the diagnosis of Alzheimer’s disease, we are carrying out “Research on AD Early Diagnosis Evaluation Model Based on Multimodal Medical Imaging Intelligent Fusion”, using AI to quickly acquire MR, and learn the relationship between MR, CT, and PET. Generate PET/CT images, and then use the trained multimodal classifier for rapid diagnosis of Alzheimer’s disease, aiming to achieve from image acquisition to final diagnosis in minutes.
AI cortical segmentation, AI Alzheimer risk monitoring, AI follow-up, AI hippocampal atrophy assessment, AI cerebral small vessel disease assessment
The combination of Medical Imaging Computing (MIC) and Computer Assisted Intervention (CAI) is the real, integrated Medical Imaging Artificial Intelligence (MICCAI).
The content of MICCAI research is also very related to the medical metaverse, not only in auxiliary diagnosis, but also in auxiliary treatment.
In terms of MIC, we have developed algorithms such as brain segmentation and registration;
In terms of CAI, we have developed intelligent surgical planning and intelligent surgical navigation; in the evaluation of postoperative efficacy, we have also developed evaluation algorithms. Combined with the new metaverse technology, immersive multi-sensing will be created for disease diagnosis, treatment planning, and intraoperative navigation.
With the continuous development of medical level and modern technology, academic research on medical AI is only valuable if it is truly implemented in clinical scenarios. I hope that more scholars can break through the boundaries and carry out more far-reaching collaboration between industry, academia, research and medicine, and jointly promote the development of cutting-edge life science and technology for the benefit of Humanity.
This article is reprinted from: https://www.leiphone.com/category/healthai/WIrTWSvZGsx0lTF7.html
This site is for inclusion only, and the copyright belongs to the original author.